Research And Improvement On Collaborative Filtering Recommendation Technology

The rapid development of Internet technology has brought great convenience topeople, especially with the advent of Web2.0technologies, people have been con-stantly creating a lot of information on the Internet. Too much information makesit hard for users to find interesting information quickly and efectively, leading tothe “information overload” problem. Currently, the search engine and personalizedrecommendation technology are two efective ways to solve this problem. However,the general search engine can provide the impersonalized information filtering onlywhen user has specific requirements, the emergence of personalized recommendationtechnology make up for the deficiency of search engine.As one of the most widely used and successful recommendation technolo-gies, collaborative filtering technology obtains the widely attention of researchers.This thesis improves the traditional collaborative filtering recommendation technolo-gy from the following two aspects: similarity calculation and rating prediction.Similarity calculation is the core of the collaborative filtering recommendationtechnology. At present commonly used similarity calculation methods mainly includethe Pearson correlation coefcient, cosine similarity. This thesis firstly analyses thecharacteristics and deficiencies of the traditional similarity calculation method, whichonly considers the numerical information of the common ratings to determine thesimilarity between two users and leads to inaccurate similarity measure. Then putforward a confidence function to measure degree of reliability of traditional similarityfrom aspect of non-numerical of common ratings. The new similarity is defined asthe product of traditional similarity and confidence function. Therefore, the new sim-ilarity calculation method combines the numerical information of the common ratingswith independent information of those values.Rating prediction is an important task of recommendation system. This thesisfirst analyzes the deficiency of the traditional rating prediction method, and then pro-poses a rating prediction method based on the interests of users. The new methodconsiders the user preferences for diferent types when predicting the unknown rat-ings, which make the predicted ratings more accurate.This article designed a series of experiments to validate the efectiveness of theabove two improved methods, the results show that the proposed similarity calculationand rating prediction method can achieve better prediction accuracy and coverage.